Introduction: The Challenge of Long-Tail Knowledge Graphs
Knowledge Graphs (KGs) represent facts as triplets $(h, r, t)$, such as (Brandon, plays_sport, baseball). While they are foundational for many AI applications, real-world KGs are often incomplete and exhibit a severe "long-tail" distribution. A few relations possess thousands of examples, while the vast majority are associated with only a handful of triplets. This data scarcity makes it difficult for traditional reasoning models to generalize effectively to infrequent relations.
Existing few-shot methods attempt to address this via meta-learning or self-supervised learning but often fail to generate high-quality negative samples, which are crucial for discriminative learning. Standard data augmentation techniques used in computer vision or natural language processing are inapplicable to KGs due to the discrete and factual nature of graph data; you cannot simply "tweak" a triplet without creating a nonsensical data point.
Methodology: The FS-KEN Framework
The proposed FS-KEN (Few-shot Knowledge Graph Reasoning by Adversarial Negative Enhancing) framework addresses these challenges through a dual-module architecture that combines adversarial training with structural graph analysis.
1. Adversarial Negative Enhancing (GAN Backbone)
The core innovation is a Policy Gradient-based Generative Adversarial Network (GAN) designed to generate "hard" negative samples. Unlike random sampling, which often produces obvious fakes (e.g., replacing "baseball" with "refrigerator"), the FS-KEN generator learns to select entities that are plausible but incorrect (e.g., replacing "baseball" with "football").
- Generator ($G$): Acts as a Reinforcement Learning (RL) agent selecting candidate entities to form negative triplets.
- Discriminator ($D$): Learns to distinguish between true triplets and the high-quality fakes produced by $G$.
- Policy Gradient: Since node selection in a graph is a discrete step, the framework uses Policy Gradient methods to allow the GAN to backpropagate gradients and improve the generator's performance.
2. Self-supervised Learning (SSL) Network
To capture relational signatures, the SSL module extracts "enclosing subgraphs" around target entities. It identifies the structural "signature" that defines a specific relation through a masking mechanism and a reconstruction loss. This ensures the model captures the most relevant relational features rather than just individual node embeddings.
Experimental Evaluation
FS-KEN was evaluated on two benchmark datasets: NELL-One (68,544 entities, 291 relations) and FB15K-237 (14,543 entities, 200 relations). The experiments focused on 3-shot scenarios, where only three samples are provided for training each relation.
Key Results
- State-of-the-Art Performance: FS-KEN outperformed 13 baseline models, including meta-learning and other self-supervised methods. On NELL-One, it achieved a 1.8% improvement in Hits@5 over the previous best model (SARF).
- Significant Gains on FB15K-237: The model showed even stronger results on FB15K-237, with Hits@5 and Hits@10 increasing by 5.16% and 3.5%, respectively.
- Efficiency: Despite the added complexity of adversarial training, inference remains fast. On FB15K-237, FS-KEN actually demonstrated faster inference times than the SARF baseline.
Conclusion and Future Directions
FS-KEN demonstrates that high-quality negative sampling is a critical bottleneck in few-shot KG reasoning. By leveraging a GAN-based approach with RL, the model effectively navigates the discrete nature of graph data to improve robustness on long-tail relations. Future research will explore broader Knowledge Graph Embedding (KGE) compatibility and the application of this framework to dynamic, evolving knowledge graphs that change over time.
Summary updated